Research map / Label-efficient and robust learning / OOD and anomaly detection
Spurious-aware OOD detection
Addresses spurious correlations and scaling in OOD detection with gradients and prototypes.
109 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2018–2026) · an approach in OOD and anomaly detection, keyword cluster “ood detection · distribution ood · distribution detection”
State History Tried Baselines Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
32 papers in 2023–24 and 46 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 15.1% to 17.8%.
Open questions the data points to
- Papers report Mahalanobis on iNaturalist (AUROC) anywhere from 89.48 to 99.54: comparisons against it depend on whose number you take.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2018 · Enhancing The Reliability of Out-of-distribution Image Detection in Neural Networks (ICLR 2018 · 634 citations)
- 2020 · Energy-based Out-of-distribution Detection (NeurIPS 2020 · 423 citations)
- 2021 · MOS: Towards Scaling Out-of-Distribution Detection for Large Semantic Space (CVPR 2021 · 159 citations)
- 2022 · ViM: Out-of-Distribution With Virtual-Logit Matching (CVPR 2022 · 273 citations)
- 2023 · Uncertainty-Aware Optimal Transport for Semantically Coherent Out-of-Distribution Detection (CVPR 2023 · 13 citations)
- 2024 · Learning Transferable Negative Prompts for Out-of-Distribution Detection (CVPR 2024 · 32 citations)
- 2025 · Detecting Out-of-Distribution Through the Lens of Neural Collapse (CVPR 2025 · 11 citations)
- 2026 · DCAC: Dynamic Class-Aware Cache Creates Stronger Out-of-Distribution Detectors (AAAI 2026 · 1 citation)
Earliest papers here
- Enhancing The Reliability of Out-of-distribution Image Detection in Neural Networks (ICLR 2018 · 634 citations)
- Out-of-Distribution Detection using Multiple Semantic Label Representations (NeurIPS 2018 · 25 citations)
- Energy-based Out-of-distribution Detection (NeurIPS 2020 · 423 citations)
What has been tried
79 papers here report 840 measured changes (a component added, removed or swapped, compared with a baseline): helped 66%, no clear effect 8%, hurt 22%, mixed 4%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Use ResNet-50 instead of ViT-B-16 as the backbone | ImageNet OOD detection, average AUC · AUC | 93.37 -> 85.70 | Hurt | Learning Transferable Negative Prompts for Out-of-Distribution Detection (CVPR 2024 · 32 citations) |
| Remove response-frequency metrics (RF+ and RF−) | CIFAR-10 common OOD · FPR95 | 3.81 -> 6.33 | Hurt | CORES: Convolutional Response-based Score for Out-of-distribution Detection (CVPR 2024 · 16 citations) |
| Aggregate all layer and class scores together instead of aggregating class-wise and taking the minimum | LOFTER · AUROC | 0.92 -> 0.83 | Hurt | Unsupervised Layer-Wise Score Aggregation for Textual OOD Detection (AAAI 2024 · 12 citations) |
| Add feature-norm filtering to NCI | ImageNet OpenOOD benchmark · Avg AUROC | 84.41 -> 88.56 | Helped | Detecting Out-of-Distribution Through the Lens of Neural Collapse (CVPR 2025 · 11 citations) |
| add informative inlier sampling | Near OOD · FPR95 | 60.57 -> 58.92 | Helped | OODD: Test-time Out-of-Distribution Detection with Dynamic Dictionary (CVPR 2025 · 9 citations) |
| Increase group size from 5 to 10 | OOD detection across the reported ID/OOD dataset pairs · AUROC | 99.2 -> 100.0; 99.7 -> 99.3 | Mixed | Learning by Erasing: Conditional Entropy Based Transferable Out-of-Distribution Detection (AAAI 2024 · 9 citations) |
| add task-adaptive proxies | near-OOD and far-OOD · AUROC | 75.18 -> 76.49; 94.85 -> 96.45 | Helped | AdaNeg: Adaptive Negative Proxy Guided OOD Detection with Vision-Language Models (NeurIPS 2024 · 8 citations) |
| add Jigsaw fake OOD data | CIFAR100, ResNet18 · AUROC | 87.43 -> 89.17 | Helped | TagFog: Textual Anchor Guidance and Fake Outlier Generation for Visual Out-of-Distribution Detection (AAAI 2024 · 7 citations) |
| Use normalized relative feature error instead of ℓ1 distance or feature ℓ1 norm alone | ImageNet-1k (average across models) · AUROC | 62.19; 40.85 -> 87.74 | Helped | CADRef: Robust Out-of-Distribution Detection via Class-Aware Decoupled Relative Feature Leveraging (CVPR 2025 · 5 citations) |
| Train for 600 epochs with RandAugment | CIFAR10 (average) · AUROC | 93.22 -> 96.76 | Helped | Exploiting Discrepancy in Feature Statistic for Out-of-Distribution Detection (AAAI 2024 · 5 citations) |
| Use CSP instead of simple adjective labels | ImageNet-1k ID, average across iNaturalist, SUN, Places, Textures · FPR95 | 24.36 -> 17.71 | Helped | Conjugated Semantic Pool Improves OOD Detection with Pre-trained Vision-Language Models (NeurIPS 2024 · 5 citations) |
| Remove classification modulation φ | ImageNet-1k OOD average, 1-shot · FPR95 | 31.09 -> 31.90 | Hurt | Self-Calibrated Tuning of Vision-Language Models for Out-of-Distribution Detection (NeurIPS 2024 · 4 citations) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| Mahalanobis | iNaturalist | AUROC | 89.48 – 99.54 | 2 papers, e.g. NeurIPS 2023 and CVPR 2022 |
What reviewers will expect
Datasets and baselines named in the 26 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- CIFAR-10 · 14 of 26 papers
- CIFAR-100 · 13 of 26 papers
- ImageNet-1k · 9 of 26 papers
- SVHN · 7 of 26 papers
- Textures · 6 of 26 papers
- iNaturalist · 6 of 26 papers
- Places365 · 6 of 26 papers
- MNIST · 5 of 26 papers
Baselines
- MSP · 14 of 26 papers
- ReAct · 12 of 26 papers
- ViM · 12 of 26 papers
- Energy · 12 of 26 papers
- ODIN · 10 of 26 papers
- Mahalanobis · 10 of 26 papers
- ASH · 8 of 26 papers
- KNN · 8 of 26 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 78 papers; from OpenAlex and the papers' affiliations).
- Tianjin University · 3 papers
- Shanghai Jiao Tong University · 3 papers
- Chinese Academy of Sciences · 3 papers
- Tencent · 2 papers
- Zhejiang University · 2 papers
- Sun Yat-sen University · 2 papers
- Peng Cheng Laboratory · 2 papers
- Guangzhou · 2 papers
- Mila · 2 papers
- CNRS · 2 papers
Use it
Download BibTeX of 39 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Provable and feature-based OOD detection (122 papers)
- Out-of-distribution generalization theory (71 papers)
- Uncertainty and misclassification detection (56 papers)
- Robust fine-tuning for OOD generalization (56 papers)
Elsewhere in the topic
- Unified multi-class anomaly detection (86 papers)
- Anomaly detection with contaminated data (64 papers)
- Deep one-class and tabular anomaly detection (50 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).